Weco AI

Weco AI builds AIDE (AI-Driven Exploration), an agent that automates machine learning engineering by treating model development as a tree-search problem. Instead of following a single trial-and-error path, AIDE organizes each code revision as a node in a search tree: when it improves a solution it branches into a new node rather than overwriting the old one, so it can compare multiple approaches and prune ones that fail or underperform. Users describe a task in natural language, and AIDE experiments on the target machine, debugging and refining candidate solutions until it converges on working code. The company reports that AIDE ranked above the median human competitor in half of the Kaggle competitions it was tested on, and that on OpenAI's MLE-Bench it earned roughly three times as many medals as the next-best agent, with a further boost when paired with reasoning models such as o1-preview. AIDE is open source, distributed as the aideml Python package and weco-cli on GitHub, making it accessible to data scientists and ML engineers who want to automate parts of their experimentation workflow. Weco AI positions itself around "research that ships" and keeping a human in the loop rather than fully removing human judgment. Pricing and any hosted or enterprise offering are not disclosed on the public site.